A curated list for Efficient Large Language Models

Model compression toolkit engineered for enhanced usability, comprehensiveness, and efficiency.

[ICML24] Pruner-Zero: Evolving Symbolic Pruning Metric from scratch for LLMs

D^2-MoE: Delta Decompression for MoE-based LLMs Compression

[ICLR 2024] Jaiswal, A., Gan, Z., Du, X., Zhang, B., Wang, Z., & Yang, Y. Compressing llms: The truth is rarely pure and never simple.

LLM Compression Benchmark

papers of llm compression

LLM Inference on AWS Lambda

[CAAI AIR'24] Minimize Quantization Output Error with Bias Compensation

QuantLRM: Quantization of Large Reasoning Models via Fine-Tuning Signals. Support Qwen, Olmo3, Llama, etc.

An implementation of the MoDeGPT LLM compression from the ICLR 2025 Conference paper: Modular Decomposition For Large Language Model Compression.

[ICLR2026] When Reasoning Meets Compression: Understanding the Effects of LLMs Compression on Large Reasoning Models. Support interpretation of Qwen, Llama, etc.

A standard PyTorch implementation of Google’s paper Language Modeling Is Compression—with no reliance on Haiku or JAX. Drawing on the original repository (https://github.com/google-deepmind/language_modeling_is_compression), this code is capable of reproducing the key results from the paper.

Token Price Estimation for LLMs

Research code for LLM Compression using Functional Algorithms, exploring stratified manifold learning, clustering, and compression techniques. Experiments span synthetic datasets (Swiss Roll, Manifold Singularities) and real-world text embeddings (DBpedia-14). The goal is to preserve semantic structure while reducing model complexity.

Knowledge-preserving SVD compression for LLMs. CF90 method: TruthfulQA +5%, 75% fact retention, importance-guided singular value selection.

Behavioral auditing toolkit for LLMs: rho-audit measures factual accuracy, bias, sycophancy, toxicity, and reasoning via teacher-forced confidence probes. SVD compression with knowledge preservation. Steering vectors for runtime behavioral control. 12-model merge audit across SLERP/TIES/DARE-TIES/Linear.

NYCU Edge AI Final Project Using SGLang